A suggested metric for cepstral ARMA-based speech classification

In this paper, we purpose a theoretical development of a metric for speech classification based on cepstral features obtained from ARMA models. Thus working with an ARMA model as a complex rational function, is possible to define a metric d(M,M´) between two stable ARMA models M, M´by means of the cepstrum coefficients of the models. This metric may be calculated algorithmically as a finite sum in the pole-zero domain. We suggest that the metric can be used in at least two circumstances: first, we might a large number of signals that come from various types of pathological sources and we wish to classify them; alternatively, we might the underlying models M i corresponding to several pathological voices and we wish to classify a voice (modeled as M, say) from one of those. In that case, we compute d(M,M i) for each i and we guess the (M i) closest to the model M.